Robust Detection of Dummy Data Attacks in Smart Grids via Transformer-Based Reinforcement Learning Under Partial Observability
针对智能电网中能模仿正常数据从而绕过现有检测的虚假数据攻击，提出一个统一检测框架，结合图谱小波变换等多维特征提取和基于Transformer的强化学习，在部分可观测条件下实现鲁棒检测，并在IEEE标准总线系统上验证了有效性。
Smart grids, as critical cyber-physical systems (CPSs) integrating advanced sensing, communication, and control technologies, face heightened vulnerability to sophisticated cyber threats. Among these emerging threats, dummy data attacks (DDAs) are becoming increasingly prominent due to their unique evasion capabilities. Unlike traditional false data injection attacks (FDIAs), DDAs optimize attack vectors to statistically mimic normal measurement data, thereby effectively bypassing existing data-driven detection mechanisms. To address this challenge, this article presents a unified detection framework for DDAs. First, a multiconstraint optimization model is constructed to simulate the generation process of stealthy DDAs. Second, a multidimensional feature extraction framework based on graph spectral wavelet transforms, topological data analysis, and centrality metrics is developed to capture subtle perturbations in voltage signals from multiple perspectives, including frequency domain, spatial structure, and topological geometry. Third, a partially observable Markov decision process (POMDP) is integrated with a transformer-based reinforcement learning (RL) architecture, enabling robust and adaptive detection under partial observability conditions. Furthermore, uncertainty estimation mechanisms are incorporated to quantify detection confidence, coupled with spatial consistency constraints to enhance decision reliability. Finally, simulation results on IEEE benchmark bus systems demonstrate the effectiveness of the proposed detection method.